NPM package does not provide an ECMAScript module
jameshfisher opened this issue · comments
Context: I want to use TensorFlow.js in the browser. I'm using rollup to bundle my app for the browser.
The TensorFlow.js source imports from seedrandom
in several places. This comes through in the ES modules distributed via npm.
To find the module seedrandom
imported by TensorFlow, I'm using the plugin @rollup/plugin-node-resolve
. This does find the node_modules/seedrandom
package.
Unfortunately, @rollup/plugin-node-resolve
does not find an ECMAScript module in the seedrandom package. So instead, it decides to bundle the seedrandom index.js
. But this is a CommonJS module, not an ES module.
As a result, rollup gives me errors like alea is not exported by node_modules/seedrandom/index.js
, and the browser gives me errors like require is not defined
.
I think the right solution is for the seedrandom package to provide an ES module. The alternative is for TensorFlow to stop importing from a package that doesn't provide an ES module.
(You might say "well you just need to use a CommonJS->ES transpiler, like @rollup/plugin-commonjs". I'm reluctant to do so, because this transformation is gross, and in general, impossible.)
@jameshfisher, did you manage to get it working? I just came across the exact same issue and followed you here from tensorflow/tfjs#4003.
Unfortunately not and I haven't had time to work on a PR :( To work around this, I'm currently using the tensorflow.js browser bundle, rather than bundling it myself.
I created a ESM module based fork called esm-seedrandom. The source is refactored to use ES6+ syntax to allow for treeshaking and to keep the minified output small and modular.
- API docs
- Demo
- Mocha-based unittests
- unittests for browsers
- unittests for NodeJS
- The NodeJS unittests also validate consistency with seedrandom for generated numbers as well as state save & restore.
If someone is motivated, we can likely work the ESM fork back into patches for seedrandom using Rollup's various JS output to get CJS, UMD, and IIFE formats without breaking backward compatibility.
@davidbau is there any chance of upstreaming the changes from @shanewholloway?